On this page
Definition
Function calling is a feature of modern language models that allows them to output a structured request to invoke a predefined function, with the calling application executing it and returning the result to the model.
Simple explanation
A language model, on its own, can only generate text — it cannot check today's weather or look up your order status. Function calling gives it a menu of tools it is allowed to ask for, such as 'get_weather' or 'lookup_order', each with a defined set of inputs.
When the model decides a tool is needed, it outputs a structured request naming the function and the arguments to use. Your application runs that function, sends the result back, and the model weaves the answer into its reply.
Why it matters
Function calling is the bridge between a model that can only talk and a system that can actually do useful, up-to-date things — check a database, send an email, run a calculation, or trigger a workflow.
It is also the foundation underneath most AI agents and tool-using assistants: without a reliable way to call functions, agents cannot interact with the outside world.
How it works
- 1Define toolsDevelopers describe available functions, their names, purposes and expected inputs.
- 2Model decidesGiven a user request, the model decides whether a function is needed and which one.
- 3Structured outputThe model returns the function name and arguments in a structured format, typically JSON.
- 4ExecutionThe application, not the model, actually runs the function.
- 5Result returnedThe function's output is sent back to the model to inform its final answer.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- OpenAI function callingLets developers connect GPT models to APIs, databases and internal tools.
- Claude tool useAnthropic's equivalent mechanism for structured tool invocation.
- ChatGPT plugins and actionsConsumer-facing feature built on top of function calling to fetch live data or trigger tasks.
Advantages
- Connects models to live, accurate data instead of relying on memorised training data.
- Enables real-world actions, not just text generation.
- Produces structured, machine-readable outputs that are easier to integrate.
- Forms the technical basis for agents and automation.
Limitations
- The model can still choose the wrong function or pass incorrect arguments.
- Every available function adds to the prompt and can increase cost.
- Requires careful validation, since executing model-suggested calls blindly is risky.
- Complex or overlapping tool sets can confuse the model's choice of which to use.
Common misunderstandings
- ClaimThe model itself executes the function.RealityThe model only requests a call; the surrounding application actually runs the code and returns the result.
- ClaimFunction calling means the model can do anything.RealityIt can only use functions that developers have explicitly defined and exposed to it.
Frequently asked questions
Is function calling the same as MCP?
They are related — MCP is a standard protocol for exposing tools, while function calling is the model capability that decides when to use them.
Can any AI model do function calling?
Most major current models support it, though the exact format and reliability vary between providers.
Is function calling safe?
It depends on what functions are exposed. Anything with real-world consequences should include validation or human approval.
Do I need to code to use function calling?
Yes, typically — it is a developer-facing feature used to build applications, not a chat-window feature.
How does function calling relate to agents?
Agents rely on function calling to take actions; it is one of the core building blocks of agentic systems.
The Tool Money Lab perspective
Function calling is unglamorous but essential — it is the plumbing that turns language models from clever text generators into components of real software.
We see the reliability of a product's function calling, more than the eloquence of its chat responses, as a genuine signal of engineering quality.
Conclusion
Function calling lets a language model ask for a specific action to be taken and use the result, connecting reasoning to real data and real systems.
It quietly underpins most of the AI products that feel genuinely useful today, from live data lookups to the agents that are starting to handle multi-step work.